HAF Model

• A Tensorflow implementation of Hierarchical Attention Flow for Multiple-Choice Reading Comprehension.
• Specially designed for oqmrc(password:duco) dataset.
• Project is based on a R-Net implementation.

Requirements

• Python = 3.6.0
• Tensorflow-gpu = 1.5.0
• spaCy >= 2.0.0
• tqdm

Getting started

Download dataset

Download dataset from oqmrc(password:duco), and unzip downloaded files.Remember to modify your data path before training.

Download Chinese word vector

• A pre-training word vector is required, for Chinese word vectors you can refer to here.
• We utilized Zhihu_QA word vectors in our experiments.

Preprocessing

python config --mode prepro

Train

python config --mode train

predict

python config --mode test

Result

Single-model accuracy on dev set is 0.7287.